CN111414997B - A Method for Battlefield Target Recognition Based on Artificial Intelligence - Google Patents

Abstract

The invention discloses a method for battlefield target identification based on artificial intelligence, which specifically comprises the following steps: step 1: image preprocessing optimization; step 2: optimizing the learning rate; setting a decline rate, and reducing the original learning rate after training a designated step length so as to prevent oscillation; step 3: multi-resolution learning and identification; step 4: non-maximum suppression. The beneficial effects are that: the method comprises the steps of carrying out recognition processing on acquired optical image data, specifically carrying out convolutional neural network training on related image data acquired on an air unmanned aerial vehicle by utilizing a deep neural network, aiming at three problems of overfitting, redundant recognition and lack of recognition accuracy existing in an original model, upgrading an algorithm model according to the characteristics of reconnaissance return optical images of the unmanned aerial vehicle, and intelligently generating a set of neural network capable of rapidly recognizing battlefield multi-type targets through image preprocessing optimization, learning rate optimization and transfer learning.

Description

A Method for Battlefield Target Recognition Based on Artificial Intelligence

technical field

The invention belongs to the fields of artificial intelligence and deep learning, and in particular relates to a method for object recognition based on artificial intelligence.

Background technique

Future warfare will inevitably be an intelligent military competition. In order to improve battlefield decision-making capabilities, the first problem to be solved is intelligent target recognition in a strong and complex electromagnetic environment. Non-intelligent, non-self-learning methods such as traditional machine learning and expert systems are difficult to deal with the problem of intelligent recognition of targets.

At present, deep learning algorithms have been used to realize the perception of the battlefield environment, and artificial intelligence technology is used to obtain timely and accurate comprehensive battlefield situation information to assist combat commanders to make quick command decisions.

Most of the research on target recognition is based on radar intelligence, but in complex electromagnetic environments, this kind of recognition will be severely limited. The image acquired based on the optical sensor is less disturbed by this. At present, image recognition algorithms based on artificial intelligence have problems such as overfitting, redundant recognition, and lack of recognition accuracy, which cannot meet the needs of battlefield target recognition.

Contents of the invention

The purpose of the present invention is to provide an artificial intelligence-based method for battlefield target recognition, which uses an improved deep learning intelligent algorithm for target recognition, and can improve the ability of reconnaissance drones to recognize battlefield targets in complex environments.

The technical scheme of the present invention is as follows: a method for battlefield target recognition based on artificial intelligence, specifically comprises the following steps:

Step 1: Image preprocessing optimization;

Step 2: Learning rate optimization;

Set a drop rate, after training the specified step size, reduce the original learning rate to prevent it from oscillating;

Step 3: Multi-resolution learning recognition;

Step 4: Non-maximum suppression.

Described step 1 comprises as follows:

(1) Generate a target frame;

(2) Image transformation optimization;

(3) Gaussian blur.

Step (1) in the described step 1 includes generating a target frame by the Transfer-Faster-RCNN model, first generating a target frame before identifying the target, and using a four-dimensional vector to represent the target frame (x, y, w, h ), where x is the abscissa of the center point of the target frame, y is the ordinate of the center point of the target frame, w is the width of the target frame, and h is the height of the target frame,

A=(Ax,Ay,Aw,Ah) (1)

G=(Gx,Gy,Gw,Gh) (2)

Among them, A is the original target box dataset, and G is the real target box dataset.

The input original window is mapped to a regression window G' that is closer to the real box G, and G' represents translation transformation:

G'x＝Ax+Aw·dx(A) (3)

G'y=Ay+Ah·dy(A) (4)

Wherein, dx(A), dy(A) represent the amount of translation, G'x represents the abscissa of the center point after the translation, and G'y represents the ordinate of the center point after the translation;

Step (2) in said step 1 comprises,

Find a transformation F of an image such that

F(Ax,Ay,Aw,Ah)=(G'x,G'y,G'w,G'h) (5)

The calculation of F can be achieved by translation and scaling:

G'w=Aw·dy(A) (6)

G'h＝Ah·dy(A) (7)

Among them, dy(A) represents the amount of translation, G'w represents the width of the zoomed target frame, and G'h represents the height of the zoomed target frame;

build objective function

Among them, φ(A) is the feature vector composed of the corresponding feature map, w ^T is the parameter to be learned; d(A) is the predicted value, (* means x, y, w, h, that is, each transformation corresponds to a The above objective function), in order to minimize the difference between the predicted value dx(A), dy(A), dw(A), dh(A) and the actual value tx, ty, tw, th, the cost function loss is as follows:

表示目标框真实的中心点，N表示特征图的数量，Aⁱ表示第i个特征图的目标框；In the formula,

Represents the real center point of the target frame, N represents the number of feature maps, and A ⁱ represents the target frame of the i-th feature map;

The function optimization objective w* is:

The step (2) in the described step 1 includes, before loading the data into the Faster-CNN model, at first different degrees of Gaussian blur and exposure processing are performed on the same picture:

Among them, p and q are the pixel positions in each RGB channel, and σ is the exposure degree coefficient.

In the described step 3, select the least cubic interpolation algorithm of image quality loss after processing, and this interpolation algorithm is expressed as follows with function W (m):

Among them, m is an independent variable, and a is an adjustment value.

Described step 4 comprises as follows:

(1) Calculate the area ratio IoU of the overlapping area of each target frame and its adjacent target frame;

(2) Compare the IoU with the threshold and change the confidence of adjacent target boxes:

Among them, s _i is the confidence degree of each target frame, and N _t is the set threshold.

The beneficial effect of the present invention is that: the present invention uses deep learning method to identify and process the acquired optical image data, specifically, utilizes deep neural network to perform convolutional neural network training on relevant image data acquired on aerial drones, aiming at the original model According to the three major problems of over-fitting, redundant recognition, and lack of recognition accuracy, the algorithm model is upgraded according to the characteristics of the optical image returned by UAV reconnaissance. Through image preprocessing optimization, learning rate optimization and transfer learning, intelligently generate a A neural network that can quickly identify multiple types of targets on the battlefield. Through optimization and transfer learning strategies, the problems of overfitting and redundant recognition are effectively solved, and the accuracy of target recognition is significantly improved.

Description of drawings

Figure 1 is the Faster-RCNN model;

Figure 2 is the YLOL V3 model;

Figure 3 is the Transfer-Faster RCNN model;

Figure 4 shows the recognition results of the YOLO v3 model;

Figure 5 shows the recognition results of the Faster-RCNN model.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

Usually, the training of a convolutional neural network requires a large amount of data, but due to the complexity and high timeliness of the actual battlefield environment, it is very difficult to collect a large amount of image information containing real battlefield targets. According to the requirements of the actual battlefield environment, the present invention introduces a transfer learning algorithm, adjusts and modifies an already trained model, and then satisfies and applies a new requirement. Since only the last single-layer fully-connected neural network is used to distinguish various types of images in the trained deep learning network model, the previous input layer and convolutional layer can be used as feature vector extraction for any image. And use the extracted feature vector as input to train a new classifier. According to the needs of the actual battlefield environment, the present invention introduces a transfer learning algorithm on the basis of the Faster-RCNN deep learning model shown in Figure 1, and establishes the Transfer-Faster RCNN model (as shown in Figure 3) by optimizing the trained model. shown), and then meet and apply to the requirements of battlefield target recognition.

A method for battlefield target recognition based on artificial intelligence provided by the present invention specifically includes the following steps:

Step 1: Image preprocessing optimization;

In one embodiment of the present invention, step 1 includes as follows:

(1) Generate target box

The target frame is generated by the Transfer-Faster-RCNN model. The Transfer-Faster-RCNN model supports inputting pictures of any size. Before recognizing the target, a target frame is first generated. The target frame is represented by a four-dimensional vector (x, y, w, h ), where x is the abscissa of the center point of the target frame, y is the ordinate of the center point of the target frame, w is the width of the target frame, and h is the height of the target frame,

A=(Ax,Ay,Aw,Ah) (1)

G=(Gx,Gy,Gw,Gh) (2)

Among them, A is the original target box dataset, and G is the real target box dataset.

The goal is to find a relationship so that the input original window is mapped to a regression window G' that is closer to the real frame G, and G' represents the translation transformation:

G'x＝Ax+Aw·dx(A) (3)

G'y=Ay+Ah·dy(A) (4)

Among them, dx(A) and dy(A) represent the amount of translation, G'x represents the abscissa of the center point after translation, and G'y represents the ordinate of the center point after translation.

(2) Image transformation optimization

Due to the mechanical uncertainty of the camera and its carrier platform, the direction and size of the acquired graphics will inevitably be shifted. Therefore, the images are scaled and rotated to different degrees, so that the system can be more accurate after loading the model. Sensitively identify the same target at different angles and sizes.

That is to find an image transformation F such that

F(Ax,Ay,Aw,Ah)=(G'x,G'y,G'w,G'h) (5)

The calculation of F can be achieved by translation and scaling:

G'w=Aw·dy(A) (6)

G'h＝Ah·dy(A) (7)

Among them, dy(A) represents the amount of translation, G'w represents the width of the zoomed target box, and G'h represents the height of the zoomed target box.

build objective function

Among them, φ(A) is the feature vector composed of the corresponding feature map, w ^T is the parameter to be learned; d(A) is the predicted value, (* means x, y, w, h, that is, each transformation corresponds to a The above objective function), in order to minimize the difference between the predicted value dx(A), dy(A), dw(A), dh(A) and the actual value tx, ty, tw, th, the cost function loss is as follows:

表示目标框真实的中心点，N表示特征图的数量，Aⁱ表示第i个特征图的目标框。In the formula,

Indicates the true center point of the target frame, N represents the number of feature maps, and A ⁱ represents the target frame of the i-th feature map.

The function optimization objective w* is:

(3) Gaussian blur

At the same time, taking into account the complex interference in the real battlefield environment, before loading the data into the Faster-CNN model, the same picture was first processed with different degrees of Gaussian blur and exposure:

Among them, p and q are the pixel positions in each RGB channel, and σ is the exposure degree coefficient.

After optimization, the photos after the above processing are loaded into the Faster-CNN model together with the original photos for training. Its advantage is that it not only increases the amount of target data obtained, but also improves the training accuracy. More importantly, after processing the pictures, it can well adapt to the influence of weather conditions and mechanical reasons in the real battlefield environment. Greatly improve the robustness of the system.

Step 2: Learning rate optimization;

In one embodiment of the present invention, step 2 includes as follows:

When training a model, the size of the learning rate will affect the training effect of the model. If the learning rate is too large, the loss rate of the training will fluctuate, and if the learning rate is too small, the training speed will be too slow. For this problem, the present invention takes the following approach:

First set a descent rate, and after training the specified step size, reduce the original learning rate to prevent it from oscillating.

Specifically, in the invention, the learning rate is reduced to 90% of the existing learning rate for every 10,000 steps of training. Secondly, after training for 100,000 steps, the training is interrupted and the learning rate is adjusted according to the loss rate. If the loss If the learning rate exceeds 30%, increase the learning rate by 50%. After adjustment, continue training with the previous training data to obtain a mature model with better effect.

Step 3: Multi-resolution learning recognition;

In one embodiment of the present invention, step 3 includes as follows:

In the verification of the Faster-CNN model, it was found that the lack of training set will cause the model to have a high misrecognition rate in the final detection, and it is difficult to meet the needs of the real battlefield environment. Therefore, the present invention proposes an optimization method for multi-resolution learning, specifically as follows:

In order to restore the lost information in the image, the present invention interpolates from the low-resolution image to generate a high-resolution image based on the model framework, obtains more image details and features, and supplies the neural network for learning.

Through comparative analysis of several classic image interpolation algorithms including nearest neighbor interpolation algorithm, bilinear interpolation method, and bi-cubic interpolation algorithm, the embodiment of the present invention selects the bi-cubic interpolation algorithm with the least image quality loss after processing, The interpolation algorithm is expressed by the function W(m) as follows:

Among them, m is an independent variable, and a is an adjustment value. While preserving the details of the original image as much as possible, we enlarge the original image so that the neural network can better acquire the features of the target in the image, thereby optimizing the quality of the training set.

In addition, when the image is recognized, interpolation and secondary recognition are performed to enlarge the target frame area, which greatly improves the accuracy of target recognition and reduces the false recognition rate. In subsequent experiments, the model optimized through this process can better recognize the camouflage camouflage in the real background, and has stronger adaptability to complex environments.

Step 4: Non-Maximum Suppression (NMS)

In one embodiment of the present invention, step 4 includes as follows:

After adopting the above model optimization, the recognition ability of the model has been greatly improved compared with the initial state, but there will be another problem - the over-fitting phenomenon of multiple recognitions, that is, the same target will be detected Multiple recognition target frames generate a large amount of redundant information, which is very unfavorable for auxiliary decision-making in the battlefield environment.

Aiming at this phenomenon, the present invention adopts a linear non-maximum value suppression algorithm to remove redundant target frames generated by the same target recognition, and retain the one with the best effect, thereby improving the recognition accuracy and reducing the false recognition rate.

The specific implementation process includes:

(1) Calculate the area ratio IoU (IntersectionoverUnion) of the overlapping area of each target frame and its adjacent target frame;

(2) Compare the IoU with the threshold and change the confidence of adjacent target boxes:

Wherein, s _i is the confidence degree of each target frame, and N _t is a set threshold, and N _t is set to 0.75 in one embodiment of the present invention. After this processing, redundant recognition target frames can be better filtered while ensuring the recognition accuracy.

In order to reflect the effectiveness of the battlefield target recognition method in the present invention, it will be verified through experiments below.

In order to better adapt to the complex real battlefield environment, only the reconnaissance drone was used for short-term target reconnaissance during the experiment. The specific target recognition process is as follows:

(1) The limited sample set obtained by the UAV is preprocessed by translation, rotation, scaling, and blurred exposure images, and the processed image is added to the original sample set to expand the number of sample sets.

(2) After preprocessing the acquired information, load the trained Faster-RCNN model (as shown in Figure 1) and YOLO v3 model (as shown in Figure 2) for training. During the training of the model, the learning rate is adjusted in time according to the real-time training loss rate to prevent it from falling into local optimization or overfitting.

When training the Faster-RCNN and YOLO v3 models, set the recognized target types to 3 types (aircraft, tanks, ships), modify the parameters in each convolutional layer according to the target type, the initial learning rate is 0.0003, and each After training for 10,000 steps, the learning rate is reduced to 95% of the existing learning rate, and the training is interrupted every 100,000 steps to adjust the learning rate by observing the change curve of the loss rate. After 200,000 steps of training, the model parameters are exported, and the model parameters are randomly tested. The training results are shown in Figure 4 and Figure 5.

(3) Perform image interpolation on the photos in the training set to increase its resolution and increase the picture information. Then, for the key areas with higher weights identified once, the resolution of the area is enlarged by interpolation, and a second identification is performed to improve accuracy and reduce misidentification.

(4) Based on the weights obtained in the previous step, if the weights of two adjacent recognition areas are higher than the set threshold, check their overlap rate. If the overlap rate is too high, it is considered a target, and the multiple recognition of one target and two targets will be reduced. Too close to the false recognition rate of these two cases.

60 photos were verified under each model, and a total of three groups of tests were carried out. The comparison results of the algorithms are shown in Table 1.

Table 1 Algorithm comparison

From the observation and analysis of the above experimental results, it is not difficult to see that under the same training conditions, YOLO v3 has a higher recognition speed with its structural advantage model of two detections, but its accuracy is slightly worse than the other two. After Transfer-FasterRCNN optimizes the original model, the recognition accuracy of the system has been greatly improved under the premise of maintaining the original recognition speed, and it can better adapt to the complex and changeable battlefield environment.

Claims (3)

1. The artificial intelligence-based method for battlefield target identification is characterized by comprising the following steps of:

step 1: image preprocessing optimization, comprising:

(1) Generating a target frame, specifically:

generating a target frame through a Transfer-fast-RCNN model, generating a target frame first before identifying a target, using four-dimensional vectors for the target frame to represent (x, y, w, h), wherein x is the abscissa of the center point of the target frame, y is the ordinate of the center point of the target frame, w is the width of the target frame,

A＝(Ax,Ay,Aw,Ah) (1)

G＝(Gx,Gy,Gw,Gh) (2)

wherein A is an original target frame data set, and G is a real target frame data set;

such that the input original window is mapped to a regression window G 'closer to the real box G, G' represents the translation transformation:

G′x＝Ax+Aw·dx(A) (3)

G′y＝Ay+Ah·dy(A) (4)

wherein dx (A) and dy (A) represent translation amounts, G 'x represents the abscissa of the translated center point, and G' y represents the ordinate of the translated center point;

(2) Image transformation optimization, specifically:

finding a transformation F of an image causes

F(Ax,Ay,Aw,Ah)＝(G′x,G′y,G′w,G′h) (5)

The calculation of F is achieved by translation and scaling:

G′w＝Aw·dy(A) (6)

G′h＝Ah·dy(A) (7)

where dy (A) represents the amount of translation, G 'w represents the scaled target frame width, and G' h represents the scaled target frame height;

construction of objective functions

Wherein phi (A) is a feature vector composed of corresponding feature graphs, w ^T The parameter to be learned, d (A) is the predicted value, and in order to make the predicted values dx (A), dy (A), dw (A), dh (A) and the true value differences tx, ty, tw, th minimum, the cost function loss is as follows:

in the method, in the process of the invention,

representing the true center point of the target frame, N represents the number of feature images, A ⁱ A target box representing an ith feature map;

the function optimization objective w is:

(3) Gaussian blur, in particular:

before loading data into the fast-CNN model, the same picture is first subjected to different degrees of Gaussian blur and exposure:

wherein, p and q are pixel point positions in each RGB channel;

step 2: optimizing the learning rate;

setting a decline rate, and reducing the original learning rate after training a designated step length so as to prevent oscillation;

step 3: multi-resolution learning and identification;

step 4: non-maximum suppression.

2. A method for battlefield target recognition based on artificial intelligence as recited in claim 1, wherein: in the step 3, a cubic interpolation algorithm is selected, and the interpolation algorithm is expressed as follows by a function W (m):

wherein m is an independent variable, and a is an adjustment value.

3. The method for battlefield target recognition based on artificial intelligence of claim 1, wherein said step 4 comprises the steps of:

(1) Calculating the area ratio IoU of the overlapping area of each target frame and the adjacent target frames;

(2) Comparing IoU to a threshold, changing the confidence of the adjacent target frame:

wherein s is _i For the confidence of each target frame, N _t Is a set threshold.

Images